Category-level Object Pose Estimation in Heavily Cluttered Scenes by Generalized Two-stage Shape Reconstructor

In this paper, we propose a method for robust estimation of the pose of an unknown object instance in an object category from a depth image, even if it is occluded. In cluttered scenes, objects are often mutually occluded, and at the same time, objects in a category often have various shapes. For estimating the object pose in such situations, we have previously proposed a Two-stage Shape Reconstructor to extract features by de-occluding the occluded region of a target object and absorbing shape variations in a category. However, the model could not be used except in a situation where the unoccluded mask of the target object is known, and only if the contour of the occluding object is expected to have a linear shape. To cope with this problem, we generalize the previous model to directly extract the feature of a de-occluded object from a depth image containing the object occluded by another object. We also propose a data augmentation method for effectively training the model. Through evaluations on large-scale virtual-world and real-world datasets, we confirm that the proposed method achieves promising results on pose estimation of an unknown occluded object from an observed depth image.